4 Approaches To Eliminate Bias In Healthcare A.I.

Historically, healthcare data has been focused on white men, and in the age of artificial intelligence (A.I.), this represents a challenge to train the algorithms to deliver results that are representative across the ethnic and gender spectrum. Given that existing data leans towards the white male subset of the population, this will inevitably lead to ‘algorithmic bias’ in healthcare. The latter term is what researchers define as the instances when the application of an algorithm does not account for inequities but may in fact exacerbate them in healthcare systems.

Indeed, researchers have found that inherent biases in data can amplify health inequities among racial minorities. We also covered the topic of A.I. bias in healthcare at The Medical Futurist. And while it is crucial to raise awareness of this aspect of smart algorithms, it is equally important to know about measures that can be undertaken to eliminate, rather than avoid, biases as A. I. increasingly become an integral part of the healthcare landscape. 

As such, this article will explore the measures that can be undertaken to make A.I. applications in healthcare more equitable.

De-biasing A.I.

Scrapping biases from healthcare A. I. is no simple task as algorithmic bias can be introduced at any point in the software’s development cycle. On top of traditionally biased health data, bias can be implemented inadvertently by the people developing the algorithms themselves or by the way features are selected and measured.

While eliminating bias across all these potential inclusion points poses a considerable challenge, researchers are not backing down to propose ways to deal with A. I. bias, or to ‘debias’ an algorithm. Here are 4 promising ways that researchers have recently investigated.

1. Consistent evaluation

Stanford University researchers highlighted in a perspective paper that challenges that lead to bias and disparity in biomedical A. I. are closely linked to the data collection and evaluation of algorithms. To ensure the reliable and fair use of A. I. in healthcare, they recommend the consistent evaluation and monitoring of these tools; as well as the collection of more diverse data. 

They suggest that developers of A. I. tools could, for example, use existing statistical tests that allow them to detect if the data used to train the algorithm is significantly different from the actual data they encounter in real-life settings. This could indicate biases due to the training data and the developers could accommodate accordingly.

2. Adopting algorithmic hygiene

When it comes to data collection, Dr. Sanjiv M. Narayan, a cardiologist and A. I. expert, recommends effective “algorithmic hygiene” as one of the best practices to keep bias out of A.I. This involves understanding the various causes of bias and ensuring that training data is representative as much as possible.

“Representative of what?” adds Dr. Narayan. “No data set can represent the entire universe of options. Thus, it is important to identify the target application and audience upfront, and then tailor the training data to that target.”

3. Applying debiasing methods

Consensus on an approach to measure or even define fairness has yet to be achieved, but measures are taken to increase the level of fairness in A.I. predictions. For this, researchers employ debiasing methods to help reduce or eliminate differences across groups or individuals per sensitive attributes. To assist developers in adapting existing debiasing methods in their work, IBM created the open-source AI Fairness 360 toolkit. In 2021, the company’s A. I. researchers applied the toolkit in real-life scenarios to successfully account for and reduce bias while allocating resources to women with postpartum depression.

4.   Ensuring transparency, privacy and regulatory oversight

Another crucial factor to ensure equity in A.I.’s decision-making is through transparency. This can take the form of cross-checking the decisions of the algorithm by humans and vice versa. In this way, they can hold each other accountable and help mitigate bias. And to ensure such transparency, the accompanying supporting infrastructure needs to be present. This encompasses the relevant technical, regulatory, economic and privacy infrastructures to deliver the required large and diverse data to transparently train algorithms. 

“Algorithmic performance changes as it is deployed with different data, different settings and different human-computer interactions. These factors could turn a beneficial tool into one that causes unintended harm, so these algorithms must continually be evaluated to eliminate the inherent and systemic inequities that exist in our healthcare system,” said Dr. Peter Embí, Associate Dean for Informatics and Health Services Research at Indiana University School of Medicine. “Therefore, it’s imperative that we continue to develop tools and capabilities to enable systematic surveillance and vigilance in the development and use of algorithms in healthcare.”

No fool-proof method

While there are more potential approaches to eliminate bias in A.I., Dr. Narayan highlights that none are foolproof. “The technology of A. I. is moving inexorably toward greater integration across all aspects of life,” he says. “As this happens, bias is more likely to occur through the compounding of complex systems but also, paradoxically, less easy to identify and prevent.”

However, this does not mean healthcare stakeholders should not strive to eliminate bias in A.I. “We wouldn’t think of treating patients with a new pharmaceutical or device without first ensuring its efficacy and safety,” said Dr. Embí. “In the same way, we must recognize that algorithms have the potential for both great benefit and harm and, therefore, require study.”

As such, in a similar way that we monitor and evaluate medicines and medical devices, we must monitor and evaluate A.I. not only for their effectiveness but on the basis of equity as well. And the methods outlined here could lead to a more equitable healthcare landscape with the assistance of A.I.

Written by Dr. Bertalan Meskó & Dr. Pranavsingh Dhunnoo